Collector reference: harness store locations and formats for ~50 agents, from tokscale's client registry

[willwashburn]

Summary

#177 item 4 asks for one collector issue per harness in burn's backlog (Gemini CLI, Copilot CLI, VS Code Copilot Chat, Qwen, Kiro, Kimi, Hermes, OpenClaw, Every Code, Droid, Pi, Amp, Roo/Kilo, Mux, Crush, OpenHands, iFlow, Cortex, Octofriend). tokscale (MIT, commit d8fd670, 2026-09-18) ships parsers for most of these and its registry (crates/tokscale-core/src/clients.rs, define_clients!) plus scanner.rs is the most complete public map of where they store transcripts. This issue records that map so each collector issue can start from a known path and format instead of re-discovering it. tokscale extracts only usage fields, so its parsers tell us where and what shape, not how to get text, tool calls and edits; each collector still needs its own characterization.

Format legend: J = JSONL transcript, JS = JSON per session/message, DB = SQLite, X = only reachable through a vendor API or a tokscale-specific sync cache (not a local transcript; not a relayhistory candidate).

Harness	Root (env override)	Files	Fmt	tokscale parser	Notes
Gemini CLI	$GEMINI_CLI_HOME else ~/.gemini	tmp/*/chats/*.json	JS	sessions/gemini.rs (1.4k lines)	project dir under tmp/
Antigravity CLI	same Gemini home	antigravity-cli/conversations/*.db	DB	sessions/antigravity_cli.rs	read directly
GitHub Copilot CLI	~/.copilot	session-store.db (usage events), otel/*.jsonl (+COPILOT_OTEL_FILE_EXPORTER_PATH), data.db	DB + J	sessions/copilot.rs (2.2k)	session-store.db needs no OTEL setup
VS Code Copilot Chat	VS Code globalStorage / workspaceStorage/*/chatSessions	chatSessions v3 with kind-1 incremental updates	JS	sessions/copilot_desktop.rs (1.6k)	see scanner.rs:1754 discover_copilot_vscode_sessions
Qwen Code	~/.qwen/projects	**/*.jsonl	J	sessions/qwen.rs	Claude-Code-like layout
Kiro	~/.kiro/sessions/cli/*.json(l); ~/.local/share/kiro-cli/data.sqlite3 (macOS ~/Library/Application Support/kiro-cli/); IDE Kiro/User/globalStorage/kiro.kiroagent	JS + DB	sessions/kiro.rs (6k)	three sources
Kimi CLI / Kimi Code	~/.kimi/sessions (wire.jsonl); ~/.kimi-code/sessions ($KIMI_CODE_HOME); desktop app-data	J	sessions/kimi.rs (1.3k)
Hermes Agent	$HERMES_HOME else ~/.hermes	state.db, profiles/*/state.db	DB	sessions/hermes.rs	scanner.rs:1061
OpenClaw	~/.openclaw/agents (+ legacy .clawdbot, .moltbot, .moldbot)	*/agent/openclaw-agent.sqlite (live), */sessions/*.jsonl (legacy) + .jsonl.deleted.*, .jsonl.reset.*, .zst, checkpoint files to exclude	DB + J	sessions/openclaw.rs (2.2k)	scanner.rs:399-432, 506-530 for the suffix rules; Codex-driven threads carry originator: "openclaw" in Codex rollouts
Droid (Factory)	~/.factory/sessions	*.settings.json + transcript	JS	sessions/droid.rs (1.2k)
Pi	~/.pi/agent/sessions	*.jsonl	J	sessions/pi.rs (1.3k)
Oh My Pi	~/.omp/agent/sessions	**/*.jsonl	J	sessions/omp.rs	Pi fork
Amp	$XDG_DATA_HOME/amp/threads	T-*.json	JS	sessions/amp.rs
Roo Code	~/.config/Code/User/globalStorage/rooveterinaryinc.roo-cline/tasks (+ ~/.vscode-server/data/...)	ui_messages.json, api_conversation_history.json	JS	sessions/roocode.rs	macOS ~/Library/Application Support/Code/..., Windows %APPDATA%\Code\...
Kilo Code (VS Code)	.../globalStorage/kilocode.kilo-code/tasks	same as Roo	JS	sessions/kilocode.rs
Kilo CLI	$XDG_DATA_HOME/kilo/kilo.db		DB	sessions/kilo.rs	OpenCode schema (opencode_schema.rs)
Cline	same globalStorage pattern, saoudrizwan.claude-dev/tasks; CLI: $CLINE_SESSION_DATA_DIR, $CLINE_DATA_DIR/sessions, $CLINE_DIR/data/sessions	ui_messages.json, api_conversation_history.json	JS	sessions/cline.rs	scanner.rs:1433-1480
Mux	~/.mux/sessions	session-usage.json	JS	sessions/mux.rs
Crush	$XDG_DATA_HOME/crush/projects.json registry → per-project data dir DB	DB	sessions/crush.rs	scanner.rs:1345-1430
Goose	$XDG_DATA_HOME/goose/sessions/sessions.db (macOS App Support, GOOSE_PATH_ROOT)		DB	sessions/goose.rs
Zed Agent	$XDG_DATA_HOME/zed/threads/threads.db (macOS ~/Library/Application Support/Zed/threads/, Win %LOCALAPPDATA%/Zed/threads/)		DB	sessions/zed.rs	hosted models only
Octofriend	~/.local/share/octofriend/sqlite.db		DB	via synthetic reattribution
Codebuff / Freebuff	$CODEBUFF_DATA_DIR else ~/.config/manicode (+-dev, -staging)	chat-messages.json	JS	sessions/codebuff.rs
Devin CLI / Desktop	$XDG_DATA_HOME/devin/cli/sessions.db; desktop ~/Library/Application Support/Devin/User/acp-events/*.ndjson (Linux ~/.config/Devin/...)	DB + NDJSON	sessions/devin.rs	ACP event stream
Junie	~/.junie/sessions/*/events.jsonl	J	sessions/junie.rs
Augment (Auggie)	~/.augment/sessions/*.json	JS	sessions/augment.rs	chat_history array per session
Cherry Studio	%APPDATA%\CherryStudio\Data\Agents\.claude\projects (macOS/Linux equivalents)	Claude Code JSONL	J	reuses claudecode.rs	a Claude Code host; PathRoot::AppData
Command Code	~/.commandcode/projects/**/*.jsonl	J	sessions/commandcode.rs	Claude-Code-like, v3 has per-request usage
ZCode	~/.zcode/cli/db/db.sqlite, legacy ~/.zcode/projects/**/*.jsonl	DB + J	sessions/zcode.rs
CodeBuddy / WorkBuddy	~/.codebuddy/projects/**/*.jsonl; ~/.workbuddy(-ai)/projects/**/*.jsonl + workbuddy.db	J / DB	sessions/codebuddy.rs, workbuddy.rs
MiMo Code	$XDG_DATA_HOME/mimocode/mimocode(-<channel>).db	DB	sessions/micode.rs	OpenCode-style channel DBs
Jcode	$JCODE_HOME else ~/.jcode/sessions/session_*.json + session_*.journal.jsonl	JS + J	sessions/jcode.rs
DeepSeek Harness	$DSH_HOME else ~/.dsh/sessions/**/session[.vN].jsonl[.zstd]	J (zstd)	sessions/dsh.rs
Prime Agent	$PRIME_AGENT_CODING_AGENT_DIR else ~/.prime/agent/sessions, session-artifacts/ (RLM children)	J	sessions/prime_agent.rs (2.6k)
Senpi / Kimchi / gajae-code / Reasonix / Hindsight / LM Studio / Unsloth / fx	see README table	mixed	one module each	niche
Cursor IDE	~/.cursor/projects/*/agent-transcripts/<id>/<id>.jsonl (relayhistory)	J	none	tokscale does not parse ~/.cursor; it pulls the dashboard API with a token from state.vscdb (tokscale-cli/src/cursor.rs:337-460). Nothing to borrow for #166
Antigravity IDE, Trae, Warp	X			vendor API sync caches only

Not in tokscale: Every Code, OpenHands, iFlow, Cortex (from burn's list) — those still need first-hand research.

Cross-cutting rules worth copying when the registry (#177) is designed:

• Roots are PathRoot::{Home, XdgData, Config, AppData, EnvVar{var, fallback_relative}} resolved once, with a use_env_roots switch for explicit --home runs (clients.rs:1-18, 100-170).
• VS Code extension stores have four spellings per OS plus ~/.vscode-server/data/User/globalStorage (scanner.rs:1433-1480).
• Several harnesses adopted OpenCode's SQLite schema (sessions/opencode_schema.rs: Kilo CLI, MiMo Code, and others), so one driver with a config struct serves them all — the same shape [G1] OpenCode: event-level parity across both storage layouts (SQLite opencode.db and legacy JSON storage/), with tokens, model/provider, parts, step-finish, compaction, parentID #168's "common normalizer" proposes.
• Archive / backup / checkpoint copies of transcripts (.jsonl.deleted.<ts>, .jsonl.reset.<ts>, .checkpoint.<uuid>.jsonl, .bak) exist for OpenClaw and must be excluded or deduplicated explicitly (scanner.rs:399-432, 506-530).

Proposed use

• Link this issue from each collector issue [G1] One-entry harness registry and out-of-tree local source plugins (absorb burn's collector backlog) #177 opens; copy the row into the issue body as the starting point.
• When a collector is implemented, its characterization fixture (per [G1] Checked-in harness fixture corpus and parser characterization tests (port burn's 40 fixtures) #163) supersedes the row.